[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [ESPResSo-devel] GPU based parallel integrate_vv()?
From: |
Rudolf Weeber |
Subject: |
Re: [ESPResSo-devel] GPU based parallel integrate_vv()? |
Date: |
Fri, 6 Jul 2018 09:53:48 +0200 |
User-agent: |
Mutt/1.5.24 (2015-08-30) |
Hi Bogdan,
On Thu, Jul 05, 2018 at 06:15:13PM +0300, Bogdan Tanygin wrote:
> Could you please advise.. Is it possible to run something similar to
> existing ESPResSo MPI based trajectories' integration but on the GPU? Just
> to be precise what I mean: the code snippet from the integrate.cpp " for (
> auto &p : local_cells.particles()) ": the "local cell" here could be within
> the CUDA block and the "for" is replaced by the threads' bundle within the
> single block.
The memory access pattern for short range interactions is too "random" to be
well-suited for the gpu, and single particle operations such as the velocity
verlet step do not make up a significant amount of the computation time.
So, we don't have plans there, and doing it would be highly non-trivial.
There are pure gpu md codes, though.
The planned change of the storage layout of particles will probably increase
short range performance significantly on the cpu For now, if the cpu
calculation time strongly exceeds the gpu calculation time, using more cores
can be a solution.
Regards, Rudolf